Cybersecurity arXiv cs.AI

Has This Checkpoint Been Abliterated? A Two-Signal Audit and Its Failure Map

A new audit method combines activation gaps and weight-recovery energy to detect if an open-weight LLM checkpoint has had its refusal mechanisms removed (abliterated).

Software Engineering arXiv cs.AI

An Exploratory Study on LLM-Generated Code and Comments in Code Repositories

A study of LLM-generated code in repositories suggests that such code is frequently found in test cases and exhibits high levels of internal cloning, with little direct association with bugs.

AI/ML arXiv cs.AI

SAB-LVLM: Significance-Aware Binarization for Large Vision-Language Models

SAB-LVLM introduces significance-aware binarization for Large Vision-Language Models to reduce memory and latency for resource-constrained deployment.

AI/ML arXiv cs.AI

Rank-Then-Act: Reward-Free Control from Frame-Order Progress

Rank-Then-Act (RTA) is a framework for learning control policies from expert videos without environment rewards by using a VLM as a progress-based ordinal scorer.

AI/ML arXiv cs.AI

Predicting Closed-Loop Performance of Latent World Models: Offline Checkpoint Selection for MPC and Model-Based RL Under Non-Markovian Rewards in LunarLander

Researchers introduce the Composite Reward Observability Fraction (CROF) to better select checkpoints for latent world models in Model-Based RL, significantly improving sample efficiency.

AI/ML arXiv cs.AI

Full Bayesian Reinforcement Learning via LF-IBIS

The LF-IBIS algorithm enables full Bayesian Reinforcement Learning without requiring an explicit likelihood function by combining Approximate Bayesian Computation and Importance Sampling.

AI/ML arXiv cs.AI

MedStreamBench: A Time-Aware Benchmark for Streaming and Proactive Medical Video Understanding

MedStreamBench is a new time-aware benchmark for medical video understanding, evaluating how models handle streaming data and proactive clinical alerting.

AI/ML arXiv cs.AI

Decentralized Stochastic Subgradient-type Methods with Communication Compression for Nonsmooth Nonconvex Optimization

A new general framework for decentralized stochastic subgradient-type methods is proposed to handle nonsmooth nonconvex optimization with communication compression.

AI/ML arXiv cs.AI

ProCal: Inference-Time Proposal Calibration for Open-Vocabulary Object Detection

ProCal is an inference-time calibration method that improves the localization quality of classification scores in open-vocabulary object detection.

Other arXiv cs.AI

AI Virtue: What is "Good" Knowledge in the Age of Artificial Intelligence?

An exploration of epistemic virtues in AI through digital humanities, questioning the nature of 'good' knowledge and generative value in the AI era.

AI/ML arXiv cs.AI

Scene-Conditioned PINN-GNN for Multipath RF Maps: Cross-Scene Generation and In-Scene Completion

A unified RF map construction framework combining PINNs and GNNs is presented to improve multipath propagation modeling in wireless communication.

AI/ML arXiv cs.AI

EPnG: Adaptive Expert Prune-and-Grow for Parameter-Efficient MoE Fine-tuning

EPnG is an adaptive prune-and-grow framework for MoE fine-tuning that reallocates LoRA capacity based on expert importance, achieving high performance with minimal parameters.

AI/ML arXiv cs.AI

Lightweight Safe Reinforcement Learning for End-to-End UAV Navigation

A lightweight safe RL framework for UAV navigation is proposed, utilizing asymmetric convolutions and a Lagrangian-based PPO algorithm for collision-risk awareness.

AI/ML arXiv cs.AI

Single-Channel EEG-Based Cognitive Load Assessment in Online Learning: A Hybrid Deep Learning Approach

A feasibility study on using single-channel EEG to assess cognitive load during online learning, including a reproducible evaluation pipeline and visualization tool.

Other Hacker News

Giant trees have no trouble pumping water to top branches

A discussion regarding the biological mechanism of how giant trees transport water to their highest branches.

Other Hacker News

Steam Controller Auto-Charge – pilot to magnetic charging puck using CV

A project pilot involving computer vision to facilitate magnetic charging for the Steam Controller.

AI/ML Hacker News

Dispersion loss counteracts embedding condensation in small language models

A technical paper discussing how dispersion loss can mitigate the condensation effect in Small Language Models (SLMs).

Software Engineering Hacker News

Leanstral 1.5: Proof Abundance for All

A discussion about Leanstral 1.5, a tool or system related to formal verification and proof abundance.

Other Hacker News

Amsterdam invented the fire department

A historical piece regarding the origin of the Amsterdam fire department.

Hardware/Chips Hacker News

The circuit that lets your brain think and see

A discussion about a circuit design that enables brain-inspired thinking and visual perception.